Arabic is not a translation target
The performance gap between English and Arabic is invisible in the vendor demo and everywhere in production. What Arabic-first engineering actually requires.
The default architecture of enterprise AI treats Arabic as a translation problem: build in English, translate at the edges, ship. It fails quietly and constantly — a customer writes in Gulf dialect and gets a fusha answer to a different question; a contract clause loses its legal force in round-trip translation; an interface mirrors left-to-right assumptions into an RTL layout nobody can scan. Each failure is small. Together they decide whether the system gets used.
Where the gap actually comes from
- Training distribution: most models saw far less Arabic than English, and far less dialect than fusha — capability drops precisely where your customers live.
- Tokenization economics: Arabic text often costs materially more tokens per sentence, which compounds into real money at platform scale.
- Code-switching reality: Gulf users mix Arabic, English, and Arabizi mid-sentence; systems tested on clean monolingual inputs shatter on real traffic.
- Evaluation blindness: teams benchmark in English and assume transfer — the single most common failure we find in inherited systems.
What Arabic-first means in practice
Arabic-first is an engineering posture, not a marketing line. It means the evaluation suite includes MSA and the dialects your users actually write, and Arabic scores gate releases with the same authority as English ones. It means model selection is driven by measured Arabic performance on your tasks. It means retrieval is tested against Arabic queries over Arabic documents, and the interface is designed right-to-left from the first wireframe rather than mirrored at the end.
A model that drops twenty points in Arabic is not twenty percent worse for this market. It is unusable — and the demo will never tell you.
For government and citizen-facing systems, this is not optional polish; Arabic is the requirement and English is the convenience. The organizations that internalize this early get an advantage that compounds: every eval case, every fine-tune, every reviewed conversation deepens an Arabic capability their competitors are still outsourcing to a translation layer.